lkml.org 
[lkml]   [2019]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC][PATCH 0/3] arm64 relaxed ABI
Date
On 19/12/2018 12:52, Dave Martin wrote:
> On Tue, Dec 18, 2018 at 05:59:38PM +0000, Catalin Marinas wrote:
>> On Tue, Dec 18, 2018 at 04:03:38PM +0100, Andrey Konovalov wrote:
>>> On Wed, Dec 12, 2018 at 4:02 PM Catalin Marinas<catalin.marinas@arm.com> wrote:
>>>> The summary of our internal discussions (mostly between kernel
>>>> developers) is that we can't properly describe a user ABI that covers
>>>> future syscalls or syscall extensions while not all syscalls accept
>>>> tagged pointers. So we tweaked the requirements slightly to only allow
>>>> tagged pointers back into the kernel *if* the originating address is
>>>> from an anonymous mmap() or below sbrk(0). This should cover some of the
>>>> ioctls or getsockopt(TCP_ZEROCOPY_RECEIVE) where the user passes a
>>>> pointer to a buffer obtained via mmap() on the device operations.
>>>>
>>>> (sorry for not being clear on what Vincenzo's proposal implies)
>>> OK, I see. So I need to make the following changes to my patchset AFAIU.
>>>
>>> 1. Make sure that we only allow tagged user addresses that originate
>>> from an anonymous mmap() or below sbrk(0). How exactly should this
>>> check be performed?
>> I don't think we should perform such checks. That's rather stating that
>> the kernel only guarantees that the tagged pointers work if they
>> originated from these memory ranges.
> I concur.
>
> Really, the kernel should do the expected thing with all "non-weird"
> memory.
>
> In lieu of a proper definition of "non-weird", I think we should have
> some lists of things that are explicitly included, and also excluded:
>
> OK:
> kernel-allocated process stack
> brk area
> MAP_ANONYMOUS | MAP_PRIVATE
> MAP_PRIVATE mappings of /dev/zero
>
> Not OK:
> MAP_SHARED
> mmaps of non-memory-like devices
> mmaps of anything that is not a regular file
> the VDSO
> ...
>
> In general, userspace can tag memory that it "owns", and we do not assume
> a transfer of ownership except in the "OK" list above. Otherwise, it's
> the kernel's memory, or the owner is simply not well defined.

Agreed on the general idea: a process should be able to pass tagged pointers at the
syscall interface, as long as they point to memory privately owned by the process. I
think it would be possible to simplify the definition of "non-weird" memory by using
only this "OK" list:
- mmap() done by the process itself, where either:
  * flags = MAP_PRIVATE | MAP_ANONYMOUS
  * flags = MAP_PRIVATE and fd refers to a regular file or a well-defined list of
device files (like /dev/zero)
- brk() done by the process itself
- Any memory mapped by the kernel in the new process's address space during execve(),
with the same restrictions as above ([vdso]/[vvar] are therefore excluded)

> I would also like to see advice for userspace developers, particularly
> things like (strawman, please challenge!):

To some extent, one could call me a userspace developer, so I'll try to help :)

> * Userspace should set tags at the point of allocation only.

Yes, tags are only supposed to be set at the point of either allocation or
deallocation/reallocation. However, allocators can in principle be nested, so an
allocator could  take a region allocated by malloc() as input and subdivide it
(changing tags in the process). That said, this suballocator must not free() that
region until all the suballocations themselves have been freed (thereby restoring the
tags initially set by malloc()).

> * If you don't know how an object was allocated, you cannot modify the
> tag, period.

Agreed, allocators that tag memory can only operate on memory with a well-defined
provenance (for instance anonymous mmap() or malloc()).

> * A single C object should be accessed using a single, fixed pointer tag
> throughout its entire lifetime.

Agreed. Allocators themselves may need to be excluded though, depending on how they
represent their managed memory.

> * Tags can be changed only when there are no outstanding pointers to
> the affected object or region that may be used to access the object
> or region (i.e., if the object were allocated from the C heap and
> is it safe to realloc() it, then it is safe to change the tag; for
> other types of allocation, analogous arguments can be applied).

Tags can only be changed at the point of deallocation/reallocation. Pointers to the
object become invalid and cannot be used after it has been deallocated; memory
tagging allows to catch such invalid usage.

> * When the kernel dereferences a pointer on userspace's behalf, it
> shall behave equivalently to userspace dereferencing the same pointer,
> including use of the same tag (where passed by userspace).
>
> * Where the pointer tag affects pointer dereference behaviour (i.e.,
> with hardware memory colouring) the kernel makes no guarantee to
> honour pointer tags correctly for every location a buffer based on a
> pointer passed by userspace to the kernel.
>
> (This means for example that for a read(fd, buf, size), we can check
> the tag for a single arbitrary location in *(char (*)[size])buf
> before passing the buffer to get_user_pages(). Hopefully this could
> be done in get_user_pages() itself rather than hunting call sites.
> For userspace, it means that you're on your own if you ask the
> kernel to operate on a buffer than spans multiple, independently-
> allocated objects, or a deliberately striped single object.)

I think both points are reasonable. It is very valuable for the kernel to access
userspace memory using the user-provided tag, because it enables kernel accesses to
be checked in the same way as user accesses, allowing to detect bugs that are
potentially hard to find. For instance, if a pointer to an object is passed to the
kernel after it has been deallocated, this is invalid and should be detected.
However, you are absolutely right that the kernel cannot *guarantee* that such a
check is carried out for the entire memory range (or in fact at all); it should be a
best-effort approach.

> * The kernel shall not extend the lifetime of user pointers in ways
> that are not clear from the specification of the syscall or
> interface to which the pointer is passed (and in any case shall not
> extend pointer lifetimes without good reason).
>
> So no clever transparent caching between syscalls, unless it _really_
> is transparent in the presence of tags.

Do you have any particular case in mind? If such caching is really valuable, it is
always possible to access the object while ignoring the tag. For sure, the
user-provided tag can only be used during the syscall handling itself, not
asynchronously later on, unless otherwise specified.

> * For purposes other than dereference, the kernel shall accept any
> legitimately tagged pointer (according to the above rules) as
> identifying the associated memory location.
>
> So, mprotect(some_page_aligned_object, ...); is valid irrespective
> of where page_aligned_object() came from. There is no implicit
> derefence by the kernel here, hence no tag check.
>
> The kernel does not guarantee to work correctly if the wrong tag
> is used, but there is not always a well-defined "right" tag, so
> we can't really guarantee to check it. So a pointer derived by
> any reasonable means by userspace has to be treated as equally
> valid.

This is a disputed point :) In my opinion, this is the the most reasonable approach.

Cheers,
Kevin

> We would need to get some cross-arch buy-in for this, otherwise core
> maintainers might just refuse to maintain the necessary guarantees.
>
>
>>> 2. Allow tagged addressed passed to memory syscalls (as long as (1) is
>>> satisfied). Do I understand correctly that this means that I need to
>>> locate all find_vma() callers outside of mm/ and fix them up as well?
>> Yes (unless anyone as a better idea or objections to this approach).
> Also, watch out for code that pokes about inside struct vma directly.
>
> I'm wondering, could we define an explicit type, say,
>
> struct user_vaddr {
> unsigned long addr;
> };
>
> to replace the unsigned longs in struct vma the mm API? This would
> turn ad-hoc (unsigned long) casts into build breaks. We could have
> an explicit conversion functions, say,
>
> struct user_vaddr __user_vaddr_unsafe(void __user *);
> void __user *__user_ptr_unsafe(struct user_vaddr);
>
> that we robotically insert in all the relevant places to mark
> unaudited code.
>
> This allows us to keep the kernel buildable, while flagging things
> that will need review. We would also need to warn the mm folks to
> reject any new code using these unsafe conversions.
>
> Of course, it would be a non-trivial effort...
>
>> BTW, I'll be off until the new year, so won't be able to follow up.
> Cheers
> ---Dave
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

\
 
 \ /
  Last update: 2019-02-11 18:29    [W:0.317 / U:0.076 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site